Audio-Native Speech Recognition with a Frozen Discrete-Diffusion Language Model
A new approach to speech-to-text that rewrites entire transcripts at once instead of word-by-word
Researchers built a speech recognition system that generates complete transcripts in parallel rather than one word at a time, using a technique called diffusion that refines rough guesses over eight steps. The system achieved 6.6% word error rate on a standard benchmark — competitive with conventional approaches — while using a clever training method to connect audio features with text generation, and works across multiple languages with a single adapter.
Parallel transcription could speed up speech recognition in real-time applications by reducing latency, since all words are refined simultaneously rather than waiting for each word to be predicted sequentially. The multilingual capability with a single adapter means the system could be more practical to deploy globally without retraining separate models for each language.